Understanding the Codification of | Vibration Orchestration and Agent Engineering – Skills Which Really Import Now

Abstract

Eighteen months after Andrej Karpathy created the term ‘vibe coding,’ the concept has already evolved into something more demanding and more professionally significant. Agent engineering – which directs AI agents to build software under structured human supervision – is becoming the automatic workflow for the professional development of software, while the random and unrevised form of vibe coding that first drew attention is increasingly understood as suitable only for prototypes. This article examines what that transition means for the skills appreciated in the labour market, why the shift from codewriting to orchestration and evaluation of AI output requires capabilities that go beyond technical programming, and what it means for the broader digital skills agenda in Malta and across Europe.

Introduction

It is rare for a concept to move from creation to expiry in less than two years. Vibe coding has managed to do something close to it – it does not disappear, but is formally replaced by its own inventor within twelve months of becoming mainstream.

On 4 February 2026, Andrej Karpathy wrote a thread on X that changed the whole industry vocabulary. One year earlier, what was later called a ‘tweet of disposable thoughts’ — and created a ‘vibe coding’. That phrase spread through Slack channels, conference keynotes, and VC pitch decks around the world.

In his one-year retrospect, Karpathy admitted that the term became viral as it caught the excitement of the moment. But it was clear: while professional workflows now automatically use LLM agents, they must be accompanied by much higher oversight and scrutiny. Its preferred term for this evolution: ‘agent engineering.’

The distinction is not merely semantic. It has direct implications for what skills employers are looking for now, what training programmes to develop, and what it means to be a digitally competent professional in an economy where AI-generated code is becoming widespread.

Why the Codification of the Vibe On its own Doesn't Kinetx Sufficiently

The limitations of random and unrevised vibe coding became apparent during 2025 as organisations started using AI-generated prototypes for production and discovered that the tools are optimized for functionality, not for the operational requirements of the real systems.

Vibe coding fails when the code moves from prototype to production. The technique exceeds design, exceeds revision, and exceeds testing — which works for demos but collapses under the weight of real users, real security requirements, and real scale. The model repeated itself in the course of 2025: It looks very good, then it comes to reality.

Amazon ordered a 90-day reset on its code deployment controls following a series of incidents in 2025, at least one linked to Amazon’s own AI encoding assistant. Dave Treadwell, Vice-President Senior of Amazon’s e-commerce services, described ‘major changes in the boom radius’ in an internal document — software updates that had spread widely because the control plans lacked safeguards. Vibration coding alone did not cause the problem. Rather, incidents reflected a broader challenge: productivity gains from AI-assisted development need to be accompanied by appropriate oversight, review discipline and operational safeguards.

Quality data has enhanced operational experience. CodeRabbit’s analysis of public demands for withdrawals reported that AI co-authored code contributions carry around 2.74 times more security vulnerabilities than human-authored contributions with the same review fund. Speed without revision discipline is not an increase in productivity. It is a liability accumulation mechanism.

What Actually Needs Agent Engineering

Agent engineering is the practice of using AI-enabled encoding agents as force multipliers under your direction, while retaining full responsibility for architecture, code quality, and engineering judgement. This means using tools such as Claude Code, OpenAI’s Codex, or GitHub Copilot to manage monotonous and tedious parts of the work – the work that, if you could remove it from your plate, would allow you to deliver more value. These tools do not replace your judgement. They extend its scope.

Changing what engineers spend their time doing is fundamental. Andrej Karpathy, who created the term vibe coding, said one year after the end of the era: “Today, programming through LLM agents is increasingly becoming an automatic workflow for professionals, except with increased oversight and scrutiny. The aim is to gain leverage from the use of agents but without any compromise on the quality of the software.’

The main knowledge is that agent engineering does not remove the need for engineering skill — but redirects it. Instead of writing a code, you design the systems, restrictions, and feedback cycles that allow the AI to write code reliably. Writing clear and precise specifications – defining what a system should do, under which constraints, with what inputs, outputs, and ways of failure – is the most critical planning competence in agent engineering. Evaluating whether the AI output actually meets those specifications is the most critical competence in the revision. None of them need to know how to write the code manually. Both need to understand what a good software should do and to be able to make the difference between a code that achieves it and a code that does not.

Labour Market Signal

The labour market has been recording this change for several months. Employers are increasingly looking for experience with AI-assisted or agent-assisted workflows, code revision and orchestration of the AI system. Anthropic launched the Anthropic Certified Architect programme in March 2026. Gartner projects that 40% of enterprises’ applications will integrate task-specific AI agents by the end of 2026, up from less than 5% at the beginning of 2025. Forrester describes the change as engineers becoming ‘AI Conductors’, focused on workflow orchestrating and architectural validation.

The credential market is following the same signal. Structured training programmes focused specifically on agent engineering workflows – writing specifications, orchestration of agents, reviewing AI output – emerged as a distinct training category separate from conventional programming instruction. The ability to be certified is not the ability to write a code. It is the ability to direct AI systems to write the code, evaluate their output, and take responsibility for what is sent.

Non-Technical Dimension

Nielsen Norman Group’s research has documented non-developers building complex systems of agent AI on intuition developed through many hours of experimentation, videos on YouTube, and community fora. Generative AI allows people’s technical ability to overcome their technical knowledge in ways that no previous technology has allowed before.

These people were applying vibe coding to design how information circulates across systems, how AI agents coordinate, and how work is done – sometimes for entire teams. As they build increasingly complex systems, they are not learning about the technical functionality under the umbrella. Instead they are developing something more amorphous – a semi-striking set of behaviours, preferences, and tendencies, formed by past hours experimenting.

The policy implication of digital skills is important: skills that make someone effective in an agent engineering context are not reliably developed through conventional training in digital skills, which tends to highlight familiarity with the tool rather than systemic thinking. Developing the ability to clearly specify, critically evaluate, and retain responsibility over AI-generated outputs is a different type of competence development – closer to professional judgement than to technical training.

What this means for Malta

Malta’s economy is already characterised by high demand from enterprises for digital and ICT capacity. The National Digital Decade Strategic Roadmap 2023-2030, published by the Government of Malta, identifies the persistent shortage of specialised ICT talent and highlights skills shortages affecting Malta’s digital economy. These shortages create particular pressure to find ways to expand the productive capacity of the existing workforce without rapid and proportional increases in the supply of formally trained ICT professionals – a supply constraint that education policy alone cannot correct quickly enough to meet the present demand.

Agent engineering, well understood and properly regulated, is one mechanism to address that discrepancy – allowing professionals without formal programming training to build and maintain software tools for their own domains, under appropriate oversight. The risk is that the enthusiasm for AI-assisted development productivity gains outweighs the governance infrastructure needed to manage its security and quality implications, a model examined in detail in the second article of this series.

The opportunity – for individuals, organisations, and Malta’s digital skills ecosystem – lies in developing the competences that are in that gap: clear specification, critical evaluation, and informed responsibility for AI-generated outputs. Those are the skills that make the difference between vibe coding that creates responsibility and agent engineering that creates value.

References

This article is based on information collected from various online sources and intended solely for informational purposes. It does not constitute professional advice or recommendations. While the information provided is accurate and true to our knowledge, we make no representations or warranties of any kind, expressed or implied, regarding the completeness, accuracy, trustworthiness, appropriateness, or availability of the subjects covered. We strongly advise readers to carry out further research and consult appropriate professionals or experts before taking any decision based on the information presented in this Article.

We waive any liability for loss or damage incurred on the basis of the information provided herein.

This document has been prepared with the help of generative AI tools for editorial support and image generation. Any suggestions generated by AI have been revised, edited, and validated by the authors, who remain fully responsible for the final content.